[leehack]

Summary

Prepares llamadart and the three Apple runtime companions at 0.19.7 for release plan #1939. Merging into main publishes core 0.19.7, then the three companions; it needs the maintainer's explicit go-ahead. Draft: not for merge, tag or publish until the maintainer says so.

Base: main f9a6045c3.

Production-readiness scope

• User-facing scope: release metadata and docs. Runtime pins unchanged since 0.19.6: native llama.cpp v0.9.0, LiteRT-LM v0.19.1-4, stable-diffusion v0.3.0-3, Web bridge v0.1.83, Web LiteRT-LM @litert-lm/[email protected].
• Changelogs (core and docs): ## Unreleased becomes ## 0.19.7: WebGPU baseline line, 4 bullets and the runtime baseline. Added a bullet documenting the BackendJsonSchemaThinkingChatGeneration interface (exported from lib/backend.dart since 0.19.5; #1907), which no released section mentions (#1935 optional item). Test-, tooling- and chat-app-only changes have no bullet. Released sections untouched; docs page keeps 0.19.x and 0.18.x.
• Companions: all three 0.19.6 -> 0.19.7 (README install snippets name core ^0.19.7). llamadart_litert_lm_flutter core constraint >=0.19.6 <0.20.0 -> >=0.19.7 <0.20.0: the floor matches the core version, which narrows in the safe direction (the verifier would also accept >=0.19.6); Package.swift pins unchanged. The other two keep >=0.19.0 <0.20.0.
• Supported platforms/paths / unsupported paths: unchanged by this PR.
• Out of scope / follow-ups: genkit_llamadart and llamadart_model_store untouched; MIGRATION.md gains a 0.19.6 -> 0.19.7 entry for the macOS x86_64 ComputeDevice.auto image-generation change. Trackers Template/parser test gaps and open parity questions from the 0.19.x audits #1933, LiteRT-LM follow-ups: e2e skip, companion version-skew warning, harness and doc fixes #1934, Test and release-tooling gaps from the 0.19.x audits #1935 and 0.19.7 audit follow-ups #1938 are closed; their leftovers are on 0.19.8 tracker 0.19.8 audit follow-ups #1950 and design issue Thinking budget: end on several sequences (Qwen3-Coder XML </think> or <tool_call>) #1949.

Completeness checklist

• Declared scope is implemented.
• Unsupported combinations: N/A, no code change.
• README, installation docs, companion READMEs and all five changelogs updated.
• Regression coverage: release gates pass.
• Validation coverage registry: N/A (no capability change).
• Security review: no secrets, tokens or signed URLs.
• Follow-ups: 0.19.8 audit follow-ups #1950 (0.19.8 tracker), Thinking budget: end on several sequences (Qwen3-Coder XML </think> or <tool_call>) #1949 (design), and genkit_llamadart: thinking budget forces a bare </think> on Qwen3-Coder XML templates #1929's genkit adoption after this release.

PR type guidance

• Release preparation: versions, install snippets, lockfiles, changelogs and the LiteRT-LM companion's core constraint. No .gitattributes change; the changelogs were edited directly, no merge or rebase.
• Feature PR / Bugfix PR / Docs-only PR: N/A.

Test Plan

Pinned Flutter 3.47.1, isolated TMPDIR, full tag history.

• dart run tool/prepare_workspace.dart (plus dart pub get in packages/llamadart_validation and website/examples/observability): locks changed only for local core/companion 0.19.6 -> 0.19.7.
• dart format --output=none --set-exit-if-changed .: 0 changed.
• dart analyze: no issues.
• dart test -p vm test/unit/tooling/: all 1,726 pass. Markdown suites from tool/ci/markdown_test_suites.txt: pass.
• Full dart test -p vm / -p chrome: left to CI (no lib/ or test change in this PR).
• coverage-lib: N/A, no lib/ change.
• Other checks:
  • verify_release_docs_versions.dart and --release-prep: exit 0, "llamadart 0.19.7, llamadart_llama_cpp_flutter 0.19.7, llamadart_litert_lm_flutter 0.19.7, llamadart_stable_diffusion_flutter 0.19.7; checkout llamadart-native v0.9.0."
  • check_webgpu_bridge_tag.dart: OK. companion_pub_preflight.dart: OK for all three companions.
  • Classifier (origin/main...HEAD): high-risk (artifactConsumer).

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
high-risk-exact-head-independent-qa	exact-head release audit	a299715	PASS	Deep audit PASS at ca8f71b8; delta re-audit PASS at a2997151. Evidence llamadart-evidence/pr-1951/ (evaluator exit 2)
release-metadata-verification	versions, constraints, changelog promotion	macOS host	PASS	--release-prep exits 0.
release-doc-version-consistency	install docs, companion bumps	macOS host	PASS	verifier, both modes, exit 0.
release-companion-pub-gate	companion pub.dev availability and constraints	macOS host	PASS (pre-merge)	companion_pub_preflight.dart OK.
genkit-release, android-release-device-pool, release-representative-smokes	N/A	N/A	N/A	Genkit untouched; no runtime-code change; smokes belong to the publish step.

Review Notes

• Independent review status: delta re-audit PASS at a2997151 (after a deep audit PASS at ca8f71b8)
• CI status / head SHA: green (36 pass, 9 skipping), a299715

High-risk regression review

• Classification: high-risk (artifactConsumer)
• Implementation task: lane-0197-release-prep
• Independent blocking QA task: fresh auditor-deep sessions (not the author). Both verifier modes, check_core_archive.py and the release-notes extraction passed; released sections are byte-identical.
• Exact head / current base: a299715 / f9a6045c3
• Production-branch deletion, bypass, or miswire proof: N/A (metadata only)
• Affected-family real-model/artifact evidence: N/A
• Explicit unavailable-family or other N/A evidence: N/A
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1951.